A Large-Scale Remote Sensing Dataset and VLM-based Algorithm for Fine-Grained Road Hierarchy Classification

The dataset construction methodology is rigorous, combining Chinese administrative standards with expert review to create aligned pixel and vector annotations. The ablation studies (Table 4) provide clear evidence that the FDE and PMSE modules offer complementary benefits: jointly integrating both improves IoU by 6.37% and F1 by 5.71% over the baseline. The geometric descriptor design (Table 5) thoughtfully encodes scale ($L$, $W$), shape ($S$, $C$), and network context ($D$, $\rho$) cues that correlate with functional road class. The comparison of VLM backbones (Table 7) is comprehensive, identifying DINOv2-ViT-B as superior to CLIP variants for grade discrimination, particularly on low-grade roads.

The framework's dependence on GPT-v for grade prediction creates a reproducibility bottleneck and potential instability due to API version changes or prompt sensitivity. The hierarchy classification accuracy (F1 64.2%, SegAcc 60.6%) remains moderate, suggesting that geometric descriptors alone may be insufficient for fine-grained discrimination without additional topological or land-use context. The authors acknowledge that "road hierarchy is inherently context-dependent" and that administrative classifications do not always align with functional roles, which introduces label ambiguity that the model does not explicitly handle. Furthermore, the evaluation on CHN6-CUG only tests binary extraction, not hierarchy classification, leaving the generalization of T-HRN largely unvalidated outside SYSU-HiRoads.

The road extraction results on CHN6-CUG demonstrate competitive performance (IoU 51.81%, F1 63.89%), surpassing RCFSNet by 5.02% IoU, though this benchmark lacks hierarchy labels and thus cannot evaluate the paper's core contribution. The ablation study in Table 6 shows that incorporating the CLIP-based VLM prior yields the most substantial gain for hierarchy classification (OA 66.8% $\rightarrow$ 71.9%), but the paper omits comparisons against non-VLM alternatives such as Random Forest or MLP classifiers using identical geometric features. Without these baselines, it is difficult to isolate whether the performance gains stem from language pre-training or merely from the multi-modal fusion architecture. The statement that RoadReasoner "surpasses state-of-the-art road extraction baselines" holds for binary masks but remains unverified for the hierarchical task due to the lack of competing methods.

The SYSU-HiRoads dataset is publicly released via Zenodo with a DOI, supporting dataset reproducibility. Training details for FORCE-Net are reasonably complete, specifying batch size 2, Adam optimizer, initial learning rate $2\times 10^{-4}$, and step decay. However, the T-HRN component relies on GPT-v without disclosure of prompt templates, temperature settings, or API versioning, creating a significant barrier to independent reproduction. The geometric discretization thresholds for converting continuous measurements into textual categories (e.g., "short/medium/long") are not specified, nor are the exact weights $w_1=0.2$, $w_2=0.8$ in the endpoint matching degree formula $D_i$ adequately justified. While the authors state that "the dataset and code will be publicly released," the dependency on proprietary LLM APIs means full reproduction requires ongoing commercial service availability.